核范数优化下的多表征挤压激励自适应网络

Multi-representation Squeeze Excitation Adaptive Network Under the Optimization of Kernel Norm

多表征自适应网络(MRAN)用于无监督学习取得了显著成效.但MRAN的特征提取只关注了域在空间结构上的联系而忽略了特征通道之间的联系,在进行无监督领域自适应(UDA)分类时,决策边界附近存在大量混淆数据的情况,当使用信息熵最小化对混淆数据进行分类时,往往会产生错误分类.针对这一问题,提出了基于批量核范数最大化的多表征挤压激励自适应网络(Multi-Representation Squeeze-Excitation Adaptation Network_Batch Kernel Norm Maximization,MRSEAN_BNM).该网络采用挤压激励注意力机制对多表征特征进行重标定,以强化重要的表征特征,采用条件最大均值差异(CMMD)拉近源域和目标域的特征分布距离,并通过最大化目标域分类输出矩阵的核范数以约束决策边界的混淆数据,达到提升域适应图像分类精度的效果.在基于公开数据集的域适应下的图像分类、可视化结果实验结果表明,MRSEAN_BNM分类精度有明显提升.

Multi-representation adaptive network(MRAN)has achieved remarkable results in unsupervised learning.However,the feature extraction of MRAN only pays attention to the relationship between the domain in the spatial structure and ignores the relationship between the feature channels.When performing unsupervised domain adaptation(UDA)classification,there is a large amount of confusing data near the decision boundary.When using information entropy minimization to classify confusing data,misclassification often occurs.To solve this problem,a Multi-Representation Squeeze-Excitation Adaptation Network_Batch Kernel Norm Maximization(MRSEAN_BNM)is proposed.The network uses the squeeze incentive attention mechanism to re-calibrate multiple characterization features to strengthen important characterization features.It uses Conditional Maximum Mean Difference(CMMD)to narrow the feature distribution distance between the source domain and the target domain,and maximizes the target domain The kernel norm of the classification output matrix is used to constrain the confusing data of the decision boundary,so as to achieve the effect of improving the accuracy of the domain to adapt to the image classification.The experimental results of image classification and visualization results based on the domain adaptation of the public data set show that the classification accuracy of MRSEAN_BNM has been significantly improved.